AI Architecture Updates: May 23, 2026

1. Uber Eats case study on InfoQ shows transformer sequence modeling and listwise ranking as the new reference architecture for real-time recommendations

Leela Kumili on InfoQ. Coverage of Uber’s revamped restaurant recommendation stack details a shift from hand-crafted statistical features to a transformer-based sequence model fed by a real-time signal layer, which collapses feature freshness from 24 hours down to seconds. The architecture pairs continuous ingestion of clicks and searches with listwise ranking that scores candidate restaurants together in a single inference pass, optimizing relative order rather than independent scores and cutting compute cost compared to pointwise approaches. To avoid the classic training-serving skew problem, Uber shares feature-extraction logic across both pipelines so the same code paths produce the embeddings during training and at request time. The takeaway for AI system designers is that near-real-time event streams, neural sequence models, and comparative ranking now form a coherent pattern for personalization at scale, displacing batch feature pipelines as the default. Source